Cross-Lingual NLI Project (English -> Hindi)

Experimenting with Zero-Shot Transfer using XLM-RoBERTa. The goal was to see if a model trained only on English NLI data could correctly classify Hindi sentences.

Creating a multilingual NLI system usually requires expensive labeled data in every target language. I wanted to see if we could skip that by leveraging cross-lingual transfer.

Results

I compared my fine-tuned XLM-RoBERTa against a standard mBERT baseline. Tested on the Hindi validation set (zero-shot).

Model	Hindi Accuracy
My Model (XLM-R)	67.2%
Baseline (mBERT)	57.1%

I also optimized it using ONNX quantization to make it faster for deployment.

• Speedup: 6x faster (24ms vs 143ms)
• Size: Reduced by ~73% (from 1GB to ~287MB)

Model Weights

Download Fine-Tuned Model (Google Drive)

Files

• train_xnli.ipynb: Main notebook. Handles data loading, training, comparison with mBERT, and ONNX conversion.
• app.py: Simple Streamlit app to try the model.
• RESEARCH_REPORT.md: My full analysis and observations.

How to Run

1. Install requirements: pip install -r requirements.txt (or just run the first cell in the notebook).
2. Run the notebook train_xnli.ipynb in Colab or locally.
3. To try the app: streamlit run app.py

What I Learned

• Shared Vocabulary Matters: XLM-R's larger vocab (250k) seems to help a lot with cross-lingual alignment compared to mBERT.
• Quantization is powerful: Did not expect a 4x size reduction with almost no accuracy drop.